AI & Machine Learning
RAG, agents, MCP, self-hosted models, classical ML
I wrote the proof of concept myself, then spent two years turning it into something that runs on a Tuesday morning without me. The demo is the easy half.
- ~60%of the AI SmartTalk product code, written by me
- 37languages answered in, on live assistants
- 2023building AI systems since — after ten years full-stack
Retrieval that survives real documents
RAG on PostgreSQL and pgvector: chunking, embeddings, re-ranking, incremental re-indexing. Vector search is never the hard part. The hard part is what happens when the source document changes at three in the morning and nobody tells the index.
Agents and tools, in both directions
LangChain and LangGraph workflows, and MCP implemented twice over — as a tool server clients call, and as an OAuth client of third-party servers. An assistant that cannot act on anything is a search bar with better manners.
Models you can host yourself
Ollama, vLLM, llama.cpp, on-premise and fully self-hosted installations. When the data is not allowed to leave the client’s infrastructure, an API key is not an architecture. AnonDocs came out of the same constraint: an MIT anonymiser that masks names, emails and identifiers before a prompt leaves the network, cited by Wired Italia.
Not only LLMs, and not only recently
At Redspher, price-prediction models in Python and TensorFlow combined with a bin-packing solver, wired straight into the ordering pipeline: 40% of road-vehicle purchasing automated. Classical ML in production, judged on the purchase orders it produced.
Stack
- Python
- pgvector
- embeddings · chunking · re-ranking
- LangChain
- LangGraph
- MCP
- Anthropic API
- OpenAI API
- evals
- Ollama
- vLLM
- llama.cpp
- TensorFlow